ADR-026: One topic vocabulary: the practice registry takes the same thirteen the reading list has

Context

ADR-020 decided the right thing and left a trap next to it.

The decision was to remove a limitation. Before it, DP-008 read the vocabulary split as a scope statement — a vision paper could be a note and not a recommendation, “not on judgement, not on evidence, but because there is no category for it to take.” ADR-020 said no: scope by the kind of claim, not by the domain the work was discovered in. That gate is wider, deliberately, and it is correct.

The trap is the alternative it rejected. Under Alternatives considered:

Widen primary_topic — add generative-modeling and vision-and-graphics to the practice seven. The obvious move, and it defeats the purpose. primary_topic is exactly-one and it is the browsing axis: a diffusion preconditioning practice tagged generative-modeling renders on the generative page and not beside the LLM normalization practices it is a sibling of.

That paragraph is why, in #124, this record’s own maintainer-facing prose said those two topics “stay out on purpose per ADR-020” — turning a decision that removed a limitation into a citation for one. If the ADR that widened the gate is the document people cite when narrowing it, the ADR needs replacing, not explaining.

The browsing argument, against the evidence

ADR-020’s objection was that a domain topic would pull cross-domain work onto its own page, away from the siblings it belongs beside. Two things now bear on that, neither available when it was written.

Nothing is displaced. Eight in-force practices are sourced to generative-modeling or vision-and-graphics papers. None of them uses either word as a secondary tag, so widening the primary group takes nothing away from any document that exists.

And they filed themselves correctly. Every one of the eight took a kind-of-claim topic:

practice	topic	sourced to
SOTA-186	data-pipeline	Segment Anything
SOTA-187	representation-and-encoding	Latent Diffusion, DDPM
SOTA-188	training-optimization	EDM, Progressive Distillation
SOTA-157	model-architecture	LLaDA
SOTA-195, SOTA-202	training-optimization	Progressive Distillation, Imagen
SOTA-201	adaptation-and-tuning	OWL-ViT
SOTA-196	analysis-and-evaluation	OWL-ViT

That is the population ADR-020 was worried about, and it is filed on the merits. The objection was really an argument against mis-tagging, and ADR-003’s one-primary rule already answers that: the topic is a choice about the kind of claim, and making a word available does not make it the right answer. A preconditioning scheme is a training claim whether or not the word “generative” is in the list.

Absence is not neutral. ADR-020 makes this argument itself, about the domain: field it withdrew — “a field that is overwhelmingly default is not neutral; a reader takes absence for a value.” The mirror holds. A practice vocabulary with no word for generative modelling tells a browsing reader that this record has no generative-modelling practices, when the truth is that it has several, filed under other words, for good reasons that the vocabulary cannot show.

Decision

The SOTA scheme takes the same thirteen topics as the LIT scheme. Add generative-modeling and vision-and-graphics to primary_topic in record/practices.d/tags.yaml and in luria.toml.

ADR-020’s actual decision is carried forward unchanged: recommendations are scoped by the kind of claim, not by the domain the work came from. This supersedes it only to reverse the rejected alternative and to stop the document being read as a boundary.

Filing rule, so the browsing objection does not materialise

Take a domain topic when the claim is about the domain as such. Take a kind-of-claim topic when the claim was merely discovered there.

• A sampler design, a guidance scheme, the diffusion-versus-autoregressive choice — those are claims about generative modelling. generative-modeling.
• A preconditioning scheme, an annotation bootstrap, a learning-rate asymmetry — those are training, data and adaptation claims that happen to have been found in image models. Their kind, as before.

If in doubt, ask what the practice would still be true of if the domain changed. If the answer is “most things”, it is a kind-of-claim practice.

Consequences

The two vocabularies become identical, which is a simplification worth naming: LIT and SOTA now differ in what a document is rather than in what it may be about.

This is not the merge ADR-003 rejected, and the direction is the whole difference. That rejection was about merging downward — forcing LIT documents into the narrow practice vocabulary, which “puts a diffusion sampler paper under training-optimization or model-architecture and makes both pages worse.” This merges upward. Nothing is forced anywhere; both schemes gain range.

The real cost, stated: a topic in an exactly-one group cannot also be a free secondary tag. Before this, a practice could carry generative-modeling as a secondary marker alongside any primary; now it must choose. No document does that today, so the cost is prospective — and it is the same cost every other primary topic already carries.

Nothing is retagged here. The eight practices above stay where they are. Two looked like candidates and neither survived a look, which is itself the evidence that the filing rule above is what does the work rather than the vocabulary:

• SOTA-187 (train the generative model in a learned compressed latent) argues for its own filing in its own body — “the transferable claim is a compute-allocation one, and it is why this is filed by the kind of claim rather than as a diffusion technique”. It stays representation-and-encoding.
• SOTA-157 (train the language model as a masked diffusion model) is a claim about how to build a language model; the diffusion part is the means. It stays model-architecture.

So the two topics added here are available and, on today’s registry, nothing takes them. That is the expected outcome and not an argument against the addition — the vocabulary exists so the record can say what a practice is about, and a word that is available and correctly unused is doing its job. Retagging a practice to demonstrate a vocabulary change is how a vocabulary change stops being about the practices.

ADR-021’s “expect a fourth” is now spent, and then some. The practice vocabulary has gone seven → ten (ADR-021) → eleven with representation-and-encoding (#114, closing a gap ADR-021 left in the same paragraph that stated the invariant it broke) → twelve with analysis-and-evaluation (ADR-027) → thirteen here. Five decisions, and the endpoint is “the same list the other scheme always had”. That is worth noticing: the vocabulary was not short of words, it was short of the words the other half of the record already used, and each decision discovered that one topic at a time.

ADR-024 is untouched. Its seams are between topics that exist and both of these are additions; if it settles in a direction that reorganises the vocabulary, this list reorganises with it.

Alternatives considered

Amend ADR-020 in place rather than supersede it. Rejected. The misreading came from a paragraph inside a document whose decision says the opposite, and editing that paragraph leaves the same shape — a widening decision containing a narrowing argument. Retirement is a status change (DP-003) and the body stays, so ADR-020’s reasoning remains readable; what changes is that a reader arrives at the current answer first.

Add the two topics and keep them out of the exactly-one group, so they stay available as free secondary markers. Rejected: it makes them a different kind of word from the other eleven, and a two-tier topic vocabulary is a worse thing to explain than either a wide one or a narrow one.